A PV prediction model based on mechanistic data-driven feature generation with temporal cross-scale alignment mechanism

Abstract

During photovoltaic (PV) power generation, the stochastic fluctuation of solar energy poses significant challenges for grid-connected systems, making accurate PV power forecasting essential for maintaining grid reliability and stability. This study proposes a PV power forecasting model that integrates mechanistic data-driven feature generation with a temporal cross-scale alignment mechanism (TCSAM). Two key features—effective irradiance and module temperature—highly correlated with power output, are derived through irradiance calculations on the tilted PV surface and heat transfer mechanisms. Various network modules extract features at different scales, capturing both slow time-varying and time-series characteristics. The model utilizes changes in features across both long-term and short-term time scales to assess their relationship with future meteorological features, identifying critical factors that significantly influence upcoming power generation. This approach enables the model to effectively detect underlying patterns and connections between past information and future outcomes. On four seasonal test sets, the model reduces RMSE by 20 %-30 % and increases R² by 2 %-3 % compared to the best baseline, highlighting its superior performance. This study offers innovative insights to enhance the accuracy and robustness of PV power forecasting, contributing to the stable operation of power grids.